DATA CONVERTER PROCESS TECHNOLOGY
ANALOG-DIGITAL CONVERSION
1. Data Converter History
2. Fundamentals of Sampled Data Systems
3. Data Converter Architectures
4. Data Converter Process Technology
4.1 Early Processes
4.2 Modern Processes
4.3 Smart Partitioning
5. Testing Data Converters
6. Interfacing to Data Converters
7. Data Converter Support Circuits
8. Data Converter Applications
9. Hardware Design Techniques
I. Index
ANALOG-DIGITAL CONVERSION
DATA CONVERTER PROCESS TECHNOLOGY
4.1 EARLY PROCESSES
CHAPTER 4
DATA CONVERTER PROCESS
TECHNOLOGY
SECTION 4.1: EARLY PROCESSES
Walt Kester
Vacuum Tube Data Converters
The vacuum tube was the first enabling technology in the development of data
converters—starting in the 1920s and continuing well into the late 1950s. As discussed in
Chapter 1 of this book, the vacuum tube was invented by Lee De Forest in 1906
(Reference 1). A figure from the patent is shown in Figure 4.1. Vacuum tubes quickly
found their way into a variety of electronic equipment, and the Bell Telephone system
began using vacuum tube amplifiers in their telephone plants as early as 1914.
Extracted from: Lee De Forest, "Device for Amplifying Feeble Electrical
Currents," U.S. Patent 841,387, Filed October 25, 1906, Issued January 15, 1907
Figure 4.1: The Invention of the Vacuum Tube: 1906
Amplifier development has always been critical to data converter development, starting
with these early vacuum tube circuits. A significant contribution was the invention of the
feedback amplifier (op amp) by Harold S. Black in 1927 (References 2, 3, 4). Vacuum
tube circuit development continued throughout World War II, and many significant
4.1
ANALOG-DIGITAL CONVERSION
contributions came from Bell Labs. For a detailed discussion of the history of op amps,
please refer to Walt Jung's book, Op Amp Applications (Reference 5).
In the 1920s, 1930s, 1940s, and 1950s, vacuum tubes were the driving force behind
practically all electronic circuits. In 1953, George A. Philbrick Researches, Inc.,
introduced the world's first commercially available op amp, known as the K2-W
(Reference 6). A photo and schematic are shown in Figure 4.2.
Figure 4.2: The K2-W Op Amp Introduced in 1953 (Courtesy of Dan Sheingold)
Pulse code modulation (PCM) was the first major driving force in the development of
early data converters, and Alec Hartley Reeves is generally credited for the invention of
PCM in 1937. (Reference 7). In his patent, he describes a vacuum tube "counting" ADC
and DAC (see Chapter 3 of this book). Data converter development continued at Bell
Labs during the 1940s, not only for use in PCM system development, but also in wartime
encryption systems.
The development of the digital computer in the late 1940s and early 1950s spurred
interest in data analysis, digital process control, etc., and generated more commercial
interest in data converters. In 1953 Bernard M. Gordon, a pioneer in the field of data
conversion, founded a company called Epsco Engineering (now Analogic, Inc.) in his
basement in Concord, MA. Gordon had previously worked on the UNIVAC computer,
and saw the need for commercial data converters. In 1954 Epsco introduced an 11-bit,
50-kSPS vacuum-tube based SAR ADC called the DATRAC. This converter, shown in
Figure 4.3, is generally credited as being the first commercial offering of such a device.
The DATRAC was offered in a 19" × 26" × 15" housing, dissipated several hundred
watts, and sold for approximately $8000.00.
While the vacuum tube DATRAC was certainly impressive for its time, solid-state
devices began to emerge during the 1950s which would eventually revolutionize the
entire field of data conversion and electronics in general.
4.2
DATA CONVERTER PROCESS TECHNOLOGY
4.1 EARLY PROCESSES
19" × 15" × 26"
150 lbs
Courtesy,
Analogic Corporation
8 Centennial Drive
Peabody, MA 01960
http://www.analogic.com
500W
$8,500.00
Figure 4.3: 1954 "DATRAC" 11-bit, 50-kSPS SAR ADC
Designed by Bernard M. Gordon at EPSCO
Solid State, Modular, and Hybrid Data Converters
Although the transistor was invented in 1947 by John Bardeen, Walter Brattain, and
William Shockley of Bell Labs (References 8, 9, 10, 11), it took nearly a decade for the
technology to find its way into commercial applications. The overall reliability of the
devices was partly responsible for this, as the first transistors were germanium, and were
limited in terms of leakage currents, general stability, maximum junction temperature,
and frequency response.
In May of 1954, Gordon Teal of Texas Instruments developed a grown-junction silicon
transistor. These transistors could operate up to 150°C, far higher than germanium.
Additional processing refinements were to improve upon the early silicon transistors, and
eventually lead a path to the invention of the first integrated circuit in 1958 by Jack Kilby
of Texas Instruments (Reference 12).
Kilby's work was paralleled by Robert Noyce at Fairchild, who also developed an IC
concept in 1959 (Reference 13). Noyce used inter-connecting metal trace layers between
transistors and resistors, while Kilby used bond wires. As might be expected from such
differences between two key inventions, so closely timed in their origination, there was
no instant concensus on the true "IC inventor." Subsequent patent fights between the two
inventor's companies persisted into the 1960s. Today, both men are recognized as IC
inventors.
In parallel with Noyce's early IC developments, Jean Hoerni (also of Fairchild
Semiconductor) had been working on means to protect and stabilize silicon diode and
4.3
ANALOG-DIGITAL CONVERSION
transistor characteristics. Until that time, the junctions of all mesa process devices were
essentially left exposed. This was a serious limitation of the mesa process. The mesa
process is so-named because the areas surrounding the central base-emitter regions are
etched away, thus leaving this area exposed on a plateau, or mesa. In practice, this factor
makes a semiconductor so constructed susceptible to contaminants, and as a result,
inherently less stable. This was the fatal flaw that Hoerni's invention addressed. Hoerni's
solution to the problem was to re-arrange the transistor geometry into a flat, or planar
surface, thus giving the new process its name (see References 14 and 15). However, the
important distinction in terms of device protection is that within the planar process the
otherwise exposed regions are left covered with silicon dioxide. This feature reduced the
device sensitivity to contaminants; making a much better, more stable transistor or IC.
With the arrangement of the device terminals on a planar surface, Hoerni's invention was
also directly amenable to the flat metal conducting traces that were intrinsic to Noyce's
IC invention. Furthermore, the planar process required no additional process steps in its
implementation, so it made the higher performance economical as well. As time has now
shown, the development of the planar process was another key semiconductor invention.
It is now widely used in production of transistors and ICs.
At a time in the early 1960s shortly after the invention of the planar process, the three key
developments had been made as summarized in Figure 4.4. They were the (silicon)
transistor itself, the IC, and the planar process. The stage was now set for important
solid-state developments in data converters. This was to take place in three stages. First,
there would be discrete transistor and modular data converters, second there would be
hybrid data converters, and thirdly, the data converter finally became a complete,
integral, dedicated IC. Of course, within these developmental stages there were
considerable improvements made to device performance. And, as with the vacuum
tube/solid-state periods, each stage overlapped the previous and/or the next one to a great
extent.
Invention of the (Germanium) transistor at Bell Labs: John
Bardeen, Walter Brattain, and William Shockley in 1947.
Silicon Transistor: Gordon Teal, Texas Instruments, 1954.
Birth of the Integrated Circuit:
Jack Kilby, Texas Instruments, 1958 (used bond wires for
interconnections).
Robert Noyce, Fairchild Semiconductor, 1959 (used
metallization for interconnections).
The Planar Process: Jean Hoerni, Fairchild Semiconductor, 1959.
Figure 4.4: Key Solid-State Developments: 1947-1959
The first solid-state data converters utilized discrete transistors, few if any ICs, and
required multiple PC boards to implement the analog and digital parts of the conversion
process. A typical example was the HS-810, 8-bit, 10-MSPS ADC introduced in 1966 by
Computer Labs, Inc. and shown in Figure 4.5. One of the PC boards from the HS-810 is
4.4
DATA CONVERTER PROCESS TECHNOLOGY
4.1 EARLY PROCESSES
shown in Figure 4.6. (Computer Labs was later acquired by Analog Devices in 1978).
The entire converter was built from discrete transistors, resistors, and capacitors, with
practically no integrated circuits. The unit was designed to fit in a 19" rack, contained all
required power supplies, dissipated over 100 W, and cost over $10,000 at the time of
introduction. Data converters such as the HS-810 were primarily used in research
applications and in early digital radar receivers.
19" RACK-MOUNTED, >100W, >$10,000
Figure 4.5: HS-810, 8-bit, 10MSPS ADC Released by
Computer Labs, Inc. in 1966
Figure 4.6: Double-Sided PC Board from HS-810 ADC
4.5
ANALOG-DIGITAL CONVERSION
By the late 1960s and early 1970s, various IC building blocks such as op amps,
comparators, and digital logic became available which allowed a considerable reduction
in parts count in ADCs and DACs. This led to the modular data converter—basically
various combinations of ICs, transistors, resistors, capacitors, etc., mounted on a small
PC board with pins, and encapsulated in a potted plastic case. The potting compound
helped to distribute the heat throughout the module, provided some degree of thermal
tracking between critical components, and made it a little more difficult for a competitor
to reverse engineer the circuit design.
A good example of an early converter module was the Analog Devices' industry-standard
ADC12QZ, 12-bit, 40-µs SAR ADC introduced in 1972 and shown in Figure 4.7. The
ADC12QZ was the first low-cost commercial general-purpose 12-bit ADC on the market.
The converter used the quad-switch ICs in conjunction with precision thin film resistor
networks for the internal DAC (the quad switch AD550 µDAC circuits are discussed in
more detail in Chapter 3 of this book).
2"×4"×0.4", 1.8W
Figure 4.7: ADC-12QZ General Purpose 12-Bit, 40-µs SAR ADC
Introduced in 1972
Another popular process for data converters that had its origins in the 1970s is the hybrid.
Hybrid circuits are typically constructed using un-encapsulated die, or "chips," such as
ICs, resistors, capacitors, etc., which are bonded to a ceramic substrate with epoxy—in
some cases, eutetically bonded. The bond pads on the various chips are connected to pads
on the substrate with wire bonds, and interconnections between devices are made with
metal paths on the substrate, similar in concept to a PC board. The metal conductor paths
are either thick film or thin film, depending upon the process and the manufacturer. For
obvious reasons, hybrid technology is often referred to as "chip-and-wire." After
assembly, the package is sealed in an inert atmosphere to prevent contamination.
Various technologies are used to construct hybrids, including thick and thin film
conductors and resistors, and the devices tend to be rather expensive compared to ICs.
The AD572 12-bit, 25-µs ADC released by Analog Devices in 1977 is an excellent
example of a hybrid and is shown in Figure 4.8. It is significant that the AD572 was the
first 12-bit hybrid ADC circuit to obtain MIL-883B approval.
4.6
DATA CONVERTER PROCESS TECHNOLOGY
4.1 EARLY PROCESSES
1.7" × 1.1" × 0.2", 0.9W
Figure 4.8: AD572 12-Bit, 25-µs ADC, 1977
The chief motivation behind modules and hybrids was to produce data converters with
speed and resolution not achievable with the early IC processes. Hybrid circuit designers
could choose from a variety of discrete PNP, NPN, and FET transistors, IC op amps, IC
comparators, IC references, IC DACs, IC logic, etc. Coupled with the ability to perform
active in-circuit laser trimming of resistors, the hybrid circuits could achieve relatively
high levels of performance compared to what was possible in ICs alone. Customers were
willing to pay premium prices for the hybrids, because that was the only way to achieve
the desired performance. Also, there was usually a period of at least several years before
the equivalent function was achievable in completely monolithic form, thereby giving a
hybrid a reasonable product life cycle.
Today, however, the situation is reversed—the speed and resolution of modern IC data
converters is generally limited by internal process-related parasitics, and these parasitics
are much smaller than could ever be achieved in an equivalent hybrid circuit. In other
words, it would be impossible to duplicate the performance of most modern IC data
converters using conventional hybrid technology. For these reasons, hybrids today serve
relatively small niche markets today, such as dc-to-dc and synchro-to-digital converters.
Note the distinction between chip-and-wire hybrids and modern multichip modules
(MCMs) which basically use surface-mount ICs and other components on small
multilayer PC boards to achieve higher levels of functionality than possible in a single
IC.
It is also important to distinguish chip-and-wire hybrids and multichip modules from
another IC packaging technology—usually referred to as compound monolithic—where
4.7
ANALOG-DIGITAL CONVERSION
two die (usually an analog IC and a digital IC) are mounted on a single lead frame,
electrically connected with wirebonds, and encapsulated in a plastic IC package.
Calibration Processes
Nearly all data converters require some calibration to ensure overall INL, DNL, gain, and
offset errors are within specified limits. For low resolution converters, the accuracy and
matching of the various circuit components may be sufficient to ensure this. For high
resolution converters (greater than 10-bits or so), methods must generally be provided to
accomplish various types of trims. The early rack-mounted and PC board data converters
generally used potentiometers and/or selected precision resistors to accomplish the
required calibration. In many cases, a precision resistor in the circuit was "padded" with a
larger parallel resistor to achieve the desired value.
Modular data converters achieved their accuracy either by using pre-trimmed ICs and
precision resistor networks as building blocks, or by manually selecting resistors prior to
potting. An interesting trim method was used in the popular DAC-12QZ—the first
modular 12-bit DAC which was introduced in 1970. It utilized thick film resistors that
were trimmed to the appropriate values by sandblasting.
Because modular data converters had to be calibrated before potting, the effects of the
thermal shifts due to potting had to be factored into the actual trim process.
Hybrid circuits generally utilized a variety of types of trimming processes, depending
upon the process and the manufacturer. Again, the use of pre-trimmed IC building blocks,
such as the AD562 or AD565 IC DAC, minimized substrate-level trimming requirements
in such products as the AD572 mentioned previously. Other popular methods included
functional laser trimming of thick or thin film resistors on the substrate. These trimmed
resistors could be in the form of deposited substrate resistors or resistor networks bonded
to the substrate. Both thick and thin film resistor technology was utilized, although thin
film resistors generally had better stability.
4.8
DATA CONVERTER PROCESS TECHNOLOGY
4.1 EARLY PROCESSES
REFERENCES:
4.1 EARLY PROCESSES
1.
Lee De Forest, "Device for Amplifying Feeble Electrical Currents," U.S. Patent 841,387, filed
October 25, 1906, issued January 15, 1907 (The triode vacuum tube, or 'Audion', the first amplifying
device).
2.
H. S. Black, "Wave Translation System," U.S. Patent 2,102,671, filed August 8, 1928, issued
December 21, 1937 (The basis of feedback amplifier systems).
3.
H. S. Black, "Stabilized Feedback Amplifiers," Bell System Technical Journal, Vol. 13, No. 1,
January 1934, pp. 1-18 (A practical summary of feedback amplifier systems).
4.
Harold S. Black, "Inventing the Negative Feedback Amplifier," IEEE Spectrum, December, 1977
(Inventor’s 50th anniversary story on the invention of the feedback amplifier).
5.
Walter G. Jung, Op Amp Applications, Analog Devices, 2002, ISBN 0-916550-26-5.
6.
Data Sheet For Model K2-W Operational Amplifier, George A. Philbrick Researches, Inc., Boston,
MA, January 1953. See also "40 Years Ago," Electronic Design, December 16, 1995, pp. 8. (The
George A. Philbrick Research dual triode K2-W, the first commercial vacuum tube op amp).
7.
Alec Harley Reeves, "Electric Signaling System," U.S. Patent 2,272,070, filed November 22, 1939,
issued February 3, 1942. Also French Patent 852,183 issued 1938, and British Patent 538,860 issued
1939. (the ground-breaking patent on PCM. Interestingly enough, the ADC and DAC proposed by
Reeves are counting types, and not successive approximation).
8.
Ian M. Ross, "The Foundation of the Silicon Age," Bell Labs Technical Journal, Vol. 2, No. 4,
Autumn 1997.
9.
C. Mark Melliar-Smith et al, "Key Steps to the Integrated Circuit," Bell Labs Technical Journal, Vol.
2, No. 4, Autumn 1997.
10. J. Bardeen, W. H. Brattain, "The Transistor, a Semi-Conductor Triode," Physical Review, Vol. 74, No.
2, July 15, 1947 pp. 230-231 (the invention of the germanium transistor).
11. W. Shockley, "The Theory of p-n Junctions in Semiconductors and p-n Junction Transistors," Bell
System Technical Journal, Vol. 28, No. 4, July 1949, pp. 435-489 (theory behind the germanium
transistor).
12. J. S. Kilby, "Invention of the Integrated Circuit," IRE Transactions on Electron Devices, Vol. ED23, No. 7, July 1976, pp. 648-654 (Kilby’s IC invention at TI).
13. Robert N. Noyce, "Semiconductor Device-and-Lead Structure," U.S. Patent 2,981,877, filed July 30,
1959, issued April 25, 1961 (Noyce’s IC invention at Fairchild).
14. Jean A. Hoerni, "Method of Manufacturing Semiconductor Devices," U.S. Patent 3,025,589, filed
May 1, 1959, issued March 20, 1962 (the planar process—a manufacturing means of protecting and
stabilizing semiconductors).
15. Jean Hoerni, "Planar Silicon Diodes and Transistors," IRE Transactions on Electron Devices, Vol. 8,
March 1961, p. 168 (technical discussion of planar processed devices).
4.9
ANALOG-DIGITAL CONVERSION
NOTES:
4.10
DATA CONVERTER PROCESS TECHNOLOGY
4.2 MODERN PROCESSES
SECTION 4.2: MODERN PROCESSES
Walt Kester, James Bryant
Bipolar Processes
The basic bipolar IC process of the 1960s was primarily optimized to yield good NPN
transistors. However, low beta, low bandwidth PNP transistors were available on the
process—the lateral PNP and the substrate PNP. Clever circuit designers were able to
use the PNPs for certain functions such as level shifting and biasing. Bob Widlar of
Fairchild Semiconductor Corporation was one of these early pioneers, and designed the
first monolithic op amp, the µA702, in 1963. Other op amps followed rapidly, including
the µA709 and the industry-standard µA741. Another Widlar design, the µA710/µA711
comparator, was introduced in 1965. These types of linear devices, coupled with the
introduction of the 7400-series TTL logic, provided some of the key building blocks for
the modular and hybrid data converters of the 1970s. For more details of the history of op
amps, please refer to Walt Jung's excellent book, Op Amp Applications (Reference 1).
Analog Devices was founded in 1965 by Ray Stata and Matt Lorber, and focused its early
efforts on precision modular amplifiers. In 1969, Analog Devices acquired Pastoriza
Electronics, then a leader in data conversion products—thereby making a solid
commitment to both data acquisition and linear technology. In 1971, Analog Devices
acquired a small IC company, Nova Devices of Wilmington, MA, and this later led to
many monolithic linear and data converter products.
Thin Film Resistor Processes
There is another process technology which does deserve special mention, since it is
crucial to the manufacture of many linear circuits and data converters requiring stable
precision resistors and the ability to perform calibrations. This is thin film resistor
technology.
Analog Devices began its efforts to develop thin film resistor technology in the early
1970s. Much effort has been spent to develop the ability to deposit these stable thin film
resistors on integrated circuit chips, and even more effort to laser trim them at the wafer
level. They have temperature coefficients of <20 ppm/°C and matching to within
0.005%. The resistors can be made to match to within 0.01% or better without laser
trimming, but to achieve this they must be relatively large—in practice if resistors must
match to better than 0.1% or 0.05%, it is more economical to laser trim them than to
design them to meet the specification without trimming.
It is interesting to note that although it is possible to make these resistors very precise
(ratiometrically), they usually have quite wide tolerances. The reason is economic—most
applications require precision matching and low temperature coefficient but do not
actually need very high absolute precision. It is possible to optimize all three, but much
less expensive to optimize two out of three—so this is what is usually done.
4.11
ANALOG-DIGITAL CONVERSION
Many of the precision resistors used in the various data converters are laser trimmed SiCr
thin film resistors, although the new submicron and non-volatile memory processes make
laser trimming unnecessary in many new data converters, where it would have been
unavoidable in earlier generations.
In summary, the bipolar process, coupled with thin film resistors and laser wafer trim
technology, led to the proliferation of IC data converters during the 1970s, 1980s, and
1990s. For example, the AD571 was the first complete monolithic 10-bit SAR ADC
designed by Paul Brokaw and was introduced in 1978. The AD571 used a bipolar process
with integrated-injection-logic (I2L) as well as thin film laser wafer trimmed resistors. I2L
geometries were made to use a set of diffusions compatible with high performance linear
transistors (Reference 2). The AD571 was soon followed by other converters, such as the
industry-standard AD574 12-bit ADC in 1980. In addition, the IC converters provided
building blocks for high performance modular and hybrid converters during the same
period.
Complementary Bipolar (CB) Processes
Although clever IC circuit designers made the best use possible of the poor quality
substrate and lateral PNP transistors available on the NPN-based bipolar processes of the
1970s and 1980s, the lack of matching high bandwidth PNP transistors definitely limited
circuit design options in linear ICs, especially high speed op amps.
In the mid 1980s, Analog Devices developed the world's first p-epi complementary
bipolar (CB) process, and the AD840-series of op amps was introduced starting in 1988.
The fts of the PNP and NPN transistors in this first-generation 36-V CB process were
approximately 700 MHz and 900 MHz, respectively. Since the introduction of the
original CB process, several generations of faster CB processes have been developed at
Analog Devices designed for even higher speeds with lower breakdowns. Descriptions of
the ground-breaking first-generation CB process can be found in References 3 and 4.
Analog Devices' CB processes all have JFETs, allowing high input impedance op amps
as well as sample-and-hold amplifiers for data converters. The dielectrically isolated
"XFCB" process provided real breakthroughs in speed and distortion performance.
Introduced in 1992, this process yields 3-GHz PNPs and 5-GHz matching NPNs. The
"XFCB 1.5" process has 5-GHz PNPs and 9-GHz NPNs. A 5-V "XFCB 2" process has
9-GHz PNPs and 16-GHz NPNs.
The XFCB process (and later generations) has been used to produce several notable highend data converters. For example, the AD9042, designed by Roy Gosser and Frank
Murden and introduced in 1995, was the first low distortion 12-bit, 41-MSPS ADC on the
market, with greater than 80-dBc SFDR over the Nyquist bandwidth. The AD9042 was
followed by several additional XFCB converters, including the AD6645 14-bit, 80-/105MSPS ADC which was introduced in 2001, with 90-dBc SFDR and 75-dB SNR. Both
the AD9042 and the AD6645 use laser wafer trimming to achieve their high level of
performance.
4.12
DATA CONVERTER PROCESS TECHNOLOGY
4.2 MODERN PROCESSES
CMOS Processes
Metal-on-silicon (MOS) devices had their origins in late 1950s and early 1960s in the
pursuit of a process tailored for digital devices. The first complementary-metal-oxidesemiconductor (CMOS) devices began to appear in the mid-1960s, and provided both Pchannel and N-channel MOS devices on the same process. CMOS offered the potential of
much higher packing density and lower power than TTL (bipolar-based) devices, and
soon became the IC process of choice for complex VLSI digital devices. The same
advances in technology that have enabled cheap, powerful, low-power consumption
processors with large memory to revolutionize mobile telephony, portable computing and
many other fields have also revolutionized data converters.
Data converter designers soon realized the advantages of using CMOS for ADCs and
DACs. As discussed in Chapter 7 of this book, CMOS switches make ideal building
blocks for data acquisition systems. In addition, CMOS offers the ability to add digital
functionality to data converters without incurring significant cost, power, and size
penalties.
In 1974, Analog Devices combined its thin film technology with CMOS to produce the
first 10-bit multiplying CMOS DAC, the AD7520, designed by Jim Cecil and Hank
Krabbe. In 1976, Analog Devices established a CMOS IC design and manufacturing
operation in Limerick, Ireland, and rapidly introduced many more general purpose
CMOS DACs and ADCs starting in the 1970s and continuing to this day.
Although CMOS is capable of making high density low power logic very efficiently and
can make excellent analog switches, it is not quite as suitable for amplifiers and voltage
references as bipolar processes. These considerations caused process technologists to
combine bipolar and CMOS processes to achieve both low power high density logic and
high accuracy low noise analog circuitry on a single chip. The resulting processes are
more complex, and therefore more expensive, than simple bipolar and CMOS processes,
but do have better mixed-signal performance. They include BiMOS processes, which are
basically bipolar processes to which CMOS structures have been added, and linear
compatible CMOS (LC²MOS or LCCMOS), which is basically CMOS with added
bipolar capability. Analog Devices' Limerick facility in Ireland began introducing data
converters, switches, and multiplexers using its own proprietary LC2MOS process in the
mid-1980s.
However, the compromises necessary to combine features mean that neither BiMOS nor
LC2MOS offers quite as good performance as its senior parent process does in its own
speciality. Thus BiMOS and LC2MOS have not replaced bipolar, complementary bipolar,
or CMOS technology, but designers now have four processes to choose from when
designing a data converter.
4.13
ANALOG-DIGITAL CONVERSION
Modern submicron CMOS technology is cheap, fast and low powered. It is also precise—
the same techniques that enable submicron features in logic and memory ICs allow us to
manufacture matched resistors and capacitors which are smaller, cheaper and more
accurately matched without subsequent trimming than has hitherto been possible, and to
make switches with lower leakage, lower on-resistance and less stray capacitance. These
advances on their own enable the manufacture of smaller, faster, cheaper and lower
powered DACs and ADCs and the integration of complex devices which would have
been too big to put on a chip a few years ago. The technology brings an additional
bonus—logic made with these processes is so small, cheap and low powered that
incorporating auto-calibration and other computational features to improve data converter
performance and accuracy is virtually free.
Cheap, reliable, non-volatile memory is another recent process innovation which
improves the performance of new generation data converters. Gain, offset and even
linearity can be adjusted after the chip has been packaged (so packaging stresses will not
affect accuracy), at a cost far lower than that of laser trimming. Many data converters
trimmed in this way are "locked" before leaving the factory so that the calibration cannot
be damaged accidentally—and so the user cannot trim them to his system's requirements.
However, the same technology does allow users to store calibration coefficients and
similar data. Other converters allow for periodic self-calibrations for gain, offset, and
even linearity errors. Various types of "fuse blowing" or "link trimming" techniques are
quite often used in the calibration process rather than more expensive thin film laser
wafer trimming.
One feature of these new submicron CMOS processes which is both a benefit and a
problem is that they have low breakdown voltage and must operate on low voltage
supplies: 0.6-µm CMOS uses 5 V, and less for the smaller geometry processes (0.35 µm
~ 3.3 V, 0.25 µm ~ 2.5 V, 0.18 µm ~ 2 V, 0.15 µm ~ 1.5 V, and 0.13 µm ~ 1 V). This
makes them virtually useless with the traditional precision analog supplies of ±15 V.
They will, however, operate accurately and at high speed on the lower supply voltages,
making them convenient for low power circuitry. However, this reduces their dynamic
range, as their fullscale output is closer (in the case of 3-V single supply circuitry 20-dB
closer) to the noise floor.
For high speed data converters, however, the reduced signal amplitude can be an
advantage, because it is generally much easier to drive low amplitude signals with low
distortion into 50-Ω or 75-Ω loads than larger amplitude signals. The optimum amplitude
for the best compromise between noise and distortion generally ranges between 1-V and
2-V peak-to-peak in communications-oriented data converters, although there are some
exceptions to this.
A brief summary of data converter processes is given in Figure 4.9.
4.14
DATA CONVERTER PROCESS TECHNOLOGY
4.2 MODERN PROCESSES
NPN-Based bipolar
NPN-Based bipolar with JFETs and LWT thin film resistors
NPN-Based Bipolar with Integrated Injection Logic (I2L) and LWT thin film
resistors
Complementary Bipolar (CB) with JFETs and LWT thin film resistors
Dielectrically Isolated Complementary Bipolar with JFETS and LWT thin
film resistors
CMOS
CMOS and LWT thin film resistors
LC2MOS and BiCMOS with LWT thin film resistors
Hybrid (chip-and-wire)
Multichip Module (MCM)
GaAs, SiGe
Figure 4.9: Data Converter Processes
Data Converter Processes and Architectures
In the last decade, CMOS has become a dominant process for data converters—replacing
more expensive bipolar laser wafer trimmed devices. Submicron CMOS has extremely
low parasitic resistance, capacitance, and inductance, and is ideal for a number of data
converter architectures, including successive approximation, Σ-∆, and pipeline. The fineline lithography techniques associated with submicron processes allow excellent
matching between capacitors in a capacitor DAC (a fundamental building block for SAR
ADCs). The internal capacitor DACs are then trimmed by adding or subtracting small
parallel capacitors using either some form of fuse blowing, link trimming, or
autocalibration routine utilizing volatile memory. The addition of analog input
multiplexers to form a complete data acquisition system is also relatively easy due to the
high quality switches and multiplexers available in CMOS.
CMOS is also the process of choice for all types of Σ-∆ ADCs and DACs, which are also
based on switched capacitor circuits. In addition, the Σ-∆ architecture is highly digitally
intensive—another reason for utilizing the packing density and low power of CMOS.
Statistical matching techniques are popular in the multibit Σ-∆ data converters as a means
for higher resolution and dynamic range without the need for trimming.
For high speed pipelined ADCs, the digital capability of CMOS is ideal to perform the
required error correction. Fully differential circuit design techniques, coupled with the
high speed switched capacitor capabilities of CMOS, produce excellent performance.
CMOS is also excellent for high speed communications DACs, as exhibited in the
Analog Devices' TxDAC family with resolutions up to 16 bits and speeds of several
hundred MHz. To illustrate the progression of DAC performance over the last decade,
4.15
ANALOG-DIGITAL CONVERSION
Figure 4.10 shows update rate moving from less than 30 MSPS in 1994 to nearly 1 GSPS
in 2004. This is primarily due to the reduction in parasitic capacitance, inductance, and
resistance associated with the smaller and smaller submicron processes. Figure 4.11
shows a similar plot for SFDR (10-MHz output signal), which has increased from 50 dBc
in 1994 to nearly 90 dBc in 2004.
10000
3000
0.35µm CMOS
DAC 1000
UPDATE
RATE
MSPS
300
2µm BiCMOS
0.25µm CMOS
100
0.6µm CMOS
30
10
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
PRODUCT INTRODUCTION DATE
Figure 4.10: High-Speed DAC Update Rate Trend
fout = 10MHz
100
0.25µm CMOS
90
0.6µm CMOS
80
SFDR
dBc
70
2µm BiCMOS
0.35µm CMOS
60
50
40
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
PRODUCT INTRODUCTION DATE
Figure 4.11: High-Speed DAC SFDR Performance Trend
4.16
2004
DATA CONVERTER PROCESS TECHNOLOGY
4.2 MODERN PROCESSES
The trend in modern data converters is to add much more digital functionality, such as
digital filtering, multiplexing, decoding, modulation, etc., and CMOS is the ideal process
for this.
As mentioned earlier, analog building blocks such as amplifiers, mixers, and voltage
references designed in CMOS cannot achieve the levels of performance attainable in
bipolar, hence the need for a process that combines bipolar with CMOS, or BiCMOS.
BiCMOS processes are more expensive, but are useful where an ADC with an extremely
high performance analog front end is required. Functions such as mixers, sample-andholds, input buffer amplifiers, and accurate voltage references can be implemented in
bipolar, while the digital portion of the data converter is CMOS.
Multichip modules offer the flexibility of combining various IC technologies to perform
functions otherwise not possible in all-monolithic parts. For instance, high performance
RF analog front ends can be tuned to match the input impedance of IC ADCs, and
thereby increase overall bandwidth. Another example is the use of digital post processing
using FPGAs to effectively increase the sampling frequency by time interleaving several
ADCs (see Chapter 8 of this book for further discussions on this topic). Modern
multichip modules are typically constructed on small low-cost PC boards using surface
mount components and offer enhanced performance in less real estate than would be
required by discrete components.
The role of Gallium Arsenide (GaAs) in modern data converters is limited to 6-bit to 8-bit
>1-GSPS flash converters and 6-bit to 10-bit DACs. These devices are high in both cost
and power and serve small niche markets.
Silicon Germanium (SiGe) offers little as a stand-alone data converter process, but
combined with CMOS, could allow the integration of RF front ends along with the data
converter function. However, these products would probably be very application specific,
as greater flexibility can probably be achieved with the devices in separate packages.
(Refer to the next section for a general discussion of the related issue of "smart
partitioning").
A brief summary of data converter processes and how they relate to various architectures
is given in Figure 4.12.
Finally, no discussion on data converter processes would be complete without touching
upon the issue of packaging. In the last decade, there has been an increase in the demand
for small, low cost, high performance, surface mount packages suitable for mass
production using automated assembly techniques. Today this is possible, primarily
because of the lower power and small die size associated with modern submicron
processes. Many devices are suitable for packages such as those shown in Figure 4.13,
which are representative of today's trends. Smaller chip-scale-packaged (CSP) devices
are available when required, and ball-grid-array (BGA) packages are useful for high pincount, high-speed devices.
4.17
ANALOG-DIGITAL CONVERSION
CMOS:
Ideal for switched capacitor SAR, Σ-∆, Pipelined
Additional digital functionality
Volatile and non-volatile trimming at package level
BiCMOS
Useful if analog front-end requires extremely high performance
Amplifiers, mixers, SHAs, highly accurate voltage references
Calibration processes
LWT, fuse blowing, link trimming, volatile and non-volatile memory,
autocalibration
Multichip Module
Multiple ADCs and DACs, analog front ends, digital post processing
GaAs
6, 8-bit GHz flash ADCs, high power and cost
SiGe
Could be useful combined with CMOS
Figure 4.12: Data Converter Processes and Architectures
SC-70
SOT-23
MSOP8
mSOIC
8-SOIC
14-SOIC
0.1 in
(ALL PACKAGES ABOVE TO SAME SCALE)
SC-70
SOT-23
OTHER PACKAGES NOT SHOWN:
BALL-GRID ARRAY (BGA)
CHIP-SCALE PACKAGES (CSP)
Figure 4.13: Examples of Modern Data Converter Packages
When power dissipation becomes significant, larger packages come equipped with builtin heatsinking "slugs" or "epads," which can be soldered directly to the PC board ground
plane to effectively dissipate the heat. The use of high speed serial interfaces is also an
important trend in reducing the total package pin count to maintain small package
profiles.
4.18
DATA CONVERTER PROCESS TECHNOLOGY
4.2 MODERN PROCESSES
REFERENCES:
4.2 MODERN PROCESSES
1.
Walter G. Jung, Op Amp Applications, Analog Devices, 2003, ISBN 0-916550-26-5. (The first
chapter on op amp history is complete with numerous references to patents, articles, etc.).
2.
A. Paul Brokaw, "A Monolithic 10-Bit A/D Using I2L and LWT Thin-Film Resistors," IEEE Journal
of Solid State Circuits, Vol. SC-13, December 1978, pp. 736-745.
3.
"Op Amps Combine Superb DC Precision and Fast Settling," Analog Dialogue, Vol. 22, No. 2, 1988
(The AD846 IC op amp, the AD840 series, and the high speed CB process used).
4.
Jerome F. Lapham, Brad W. Scharf, "Integrated Circuit with Complementary Junction-Isolated
Transistors and Method of Making Same," U.S. Patent 4,969,823, filed May 5, 1988, issued Nov. 13,
1990. (Design of the ADI CB IC process).
ACKNOWLEDGEMENTS
Thanks are due to Doug Mercer and Dave Robertson of Analog Devices who provided valuable insights
regarding modern data converter IC processes and their relationship to the various trends and architectures.
James Bryant provided some of the process-related descriptions, and Walt Jung's Op Amp Applications
book was the primary source for semiconductor history.
4.19
ANALOG-DIGITAL CONVERSION
NOTES:
4.20
DATA CONVERTER PROCESS TECHNOLOGY
4.3 SMART PARTITIONING
SECTION 4.3
SMART PARTITIONING
Dave Robertson, Martin Kessler
When Complete Integration Isn't the Optimal Solution
For 30 years, the main path to "smaller, faster, better, cheaper" electronic devices has
been through putting more and more of a given system onto a single chip. Large rewards
were reaped by those companies that could overcome the various technologic barriers to
integration, providing more functionality on a single chip. But as we enter the very deep
submicron age, we are approaching some important physical limitations that will change
the cost and performance tradeoffs that designers have traditionally made.
As we approach the limits of practical reduction in feature size, it will increasingly turn
out that a two-chip design will be smaller, faster, better, cheaper than a single, integrated
solution. The key in these cases will be selecting the boundary between these chips.
Although high levels of integration will continue to be a feature of the most advanced
systems, reaching the optimum in cost and performance will no longer be a simple case
of steadily increasing integration. Rather, progress will be measured by changes in circuit
partitioning that enable system improvements.
There are several examples of this partitioning already evident today. For instance, large
amounts of memory are generally cheaper to implement as a separate DRAM chip than to
embed into a microprocessor. It is important to note that as integration barriers emerge,
we will not step back to the days where the design model was "analog on one chip, digital
on another, memory on a third." Chip partitioning will be done along boundaries that
optimize the flow of signal information, as well as augment the intellectual property
strengths of the chip providers. It will not be a simple case of "dis-integration." Instead,
the best systems will reflect carefully considered integration, facilitating a "smart
partitioning."
What makes one partitioning "smarter" than another? There are several important factors
to consider:
• Supply Voltage—Each advance in lithography brings with it a reduced supply
voltage. While this generally helps to lower the power in digital circuits, a lower supply
voltage can actually cause power dissipation to increase for high-performance analog
circuits. Lower supply voltages also force the use of smaller signal swings, making it
difficult to maintain good signal-to-noise ratios. Many systems will look to implement
critical analog functions on technologies that support a higher supply voltage as is being
done today in cable modem line drivers.
• Pin Count—This still drives package/assembly cost as well as board area, so it is
desirable to partition systems in a way that minimizes the number of chip-to-chip
interconnections. For example, simple digital-to-analog converter (DAC) functions can
still be best integrated onto the digital chip if it allows a single analog output pin rather
than a full 12-line digital bus.
4.21
ANALOG-DIGITAL CONVERSION
• Interface Bandwidth—A digital bus running at 500 MHz dissipates more power and
generates more EMI than a digital bus running at 5 MHz. Wherever possible, the system
should be partitioned across buses running at modest rates; in some cases, the data flow
can even be carried as a serial bus, thereby also saving pins. Often, this means putting a
large amount of digital processing on an otherwise "analog" chip. Examples of this
include decimation filters on analog-to-digital-converters (ADCs), interpolation filters on
DACs, and Direct Digital Synthesis with integrated DACs. Low voltage differential
signaling (LVDS) can be used for high-speed interfaces (>200 MHz). LVDS provides
better signal integrity and lower power dissipation at higher frequencies than a standard
CMOS interface but doubles the pin count due to its differential nature.
• Testability and Yield—Some levels of integration are technically possible, but a poor
choice from a manufacturability perspective. Integrating a finicky function with yield
issues onto a very large chip means one is forced to throw the entire large chip away each
time the function fails a test-which can be very expensive. It is far more cost effective to
segregate the function that is subject to yield fallout.
• External Components—When considering integration, it is important to factor in not
only the ICs, but also the external passive components (capacitors, inductors, SAW
filters, etc.). In many cases, an innovative architecture coupled with smart partitioning
can provide significant savings in external components, leading to much smaller formfactor and manufacturing costs. One example is illustrated in Figure 4.14 with ADI's
OTHELLO® direct-conversion chipset. The multimode cellular handset chipset combines
circuit innovation with system understanding and smart partitioning to create a
breakthrough in form factor, performance and power saving.
2.00"
1.25"
Figure 4.14: Othello Direct Conversion Radio
4.22
DATA CONVERTER PROCESS TECHNOLOGY
4.3 SMART PARTITIONING
• Flexibility—For the highest volume applications, a completely optimized solution is
generally provided in the form of full-custom ASICs. However, the vast majority of
applications never reach the run rate that justifies a fully committed integrated circuit
solution. In this case, the designer will look for the highest levels of integration and
performance available-often leveraging neighboring high-volume applications—and will
fill in around these with FPGA or other programmable solutions to customize to the
application. Examples include the use of TV tuners in cable modem boxes and the use of
cell phone handset chipsets in some low-end base stations.
• Cost—As CMOS fabrication processes move to finer geometries, digital circuits
shrink dramatically and become more cost efficient despite higher silicon wafer costs (see
Figure 4.15). Analog circuitry however, as illustrated in Figure 4.16, does not shrink as
significantly when migrating to finer process geometries. It may in fact even grow in size
to maintain performance. At finer process geometries, the overall digital per-function cost
decreases while the overall analog per-function cost increases. Moreover, chips that
integrate digital with analog functions may experience significantly higher yield losses
than pure digital chips.
DIGITAL ICs
As CMOS fab processes
move to finer geometries
Digital circuits shrink
dramatically
More features can be
packed onto the same
size die
The cost per unit area
of silicon goes up
DIE SIZE
SILICON COST/AREA
FUNCTION COST
0.6µm
Overall, per function cost
decreases because circuit sizes
shrink faster than silicon costs
increase
0.35µm
0.25µm
0.18µm
0.13µm
PROCESS GEOMETRY
Figure 4.15: Fab Process Geometry Effects on Cost of Digital ICs
Smart partitioning separates analog and mixed-signal circuits from pure digital circuitry
for cost optimization (see Figure 4.17). An example of such a partitioning is
demonstrated in Figure 4.18 with a set-top-box chip-set that can also provide cablemodem functionality. The high-density digital ASIC is separated from the analog or
mixed-signal components. All ADCs and DACs are integrated with front-end digital
functions like interpolator, DDS and modulator into a single, mixed-signal front-end, the
AD9877/AD9879 from Analog Devices.
4.23
ANALOG-DIGITAL CONVERSION
ANALOG ICs
As CMOS processes shrink
below 0.25µm
Analog circuitry does not
shrink significantly
It may in fact grow to
maintain performance
Issues of supporting
signal dynamic range
arise with decreasing
supply voltages
Designing high performance
analog circuits gets harder
and takes longer
Overall, per function cost
increases!!
DIE SIZE
SILICON COST/AREA
FUNCTION COST
0.6µm
0.35µm
0.25µm
0.18µm
0.13µm
PROCESS GEOMETRY
Figure 4.16: Fab Process Geometry Has a Different Effect on Analog Cost
Traditional Partitioning—
Digital Chip, Integrated
Mixed Signal Front End (MxFE):
Analog
MxFE
0.35µm
Digital
New Partitioning Trend—
Digital Customer ASIC, Separate
Mixed Signal Front End (MxFE):
0.18µm !
Analog
MxFE
Digital
0.35µm
NOW
$$$
0.18µm
NOW
SOON
Analog
Digital MxFE
$$
0.13µm !
SOON
Digital
Analog
MxFE
Better Overall Yield
More Cost Effective
Optimized Performance
Figure 4.17: Smart Analog/Digital Partitioning
4.24
0.18µm
DATA CONVERTER PROCESS TECHNOLOGY
4.3 SMART PARTITIONING
Cable Modem Digital ASIC
QAM
Mapping
Tx Filter
Burst Gen.
FEC
Coder
M
A
C
CTRL
Interpolator
IQ
MOD
DAC
ADC
DAC
FEC
Decoder
QAM
Demod
RISC
CTRL
NTSC / PAL Decoder
MPEG-2 Decoder
USB/
Ethernet
Video Encoder
Audio/Video CODEC
Coax
Cable
AD832x
ADC
Out-of-Band
QPSK-Demod
Diplexer
LPF
ADC
AD9877/
AD9879
BPF
DAC
T
U
N
E
R
T
U
N
E
R
BPF
ADC
Clamp
AD18xx
ADV71xx
Figure 4.18: AD9877 / AD9879 Set-Top Box Mixed Signal Front End (MxFE™)
• Performance—ADI's TxDAC® family was launched several years ago, ushering in a
new generation of CMOS digital-to-analog converters with exceptional dynamic
performance suitable for communications applications. While the product family has
three generations of stand-alone converters, it also includes several converters that take
advantage of the fine-line CMOS process by integrating digital interpolation filters.
These filters take the input data word stream and insert additional sample words that are
created by on-chip digital FIR filters. The AD9777 (Figure 4.19) features a few
interesting dimensions of smart partitioning.
AD9777
16
Digital
ASIC/
FPGA
2
2
AD8346
2
DAC
16
DATA
DEMUX/
LATCH
SELECTABLE
INTERPOLATION FILTERS
2
2
GAIN/OFFSET
CONTROL
2
DAC
PLL CLOCK MULTIPLIER
Figure 4.19: AD9777 TxDAC and AD8346 Quadrature Modulator
4.25
ANALOG-DIGITAL CONVERSION
For example, putting both DACs on the same piece of silicon significantly improves the
matching performance, which is critical for quadrature balance in many communications
applications. Furthermore, it allows a complex digital upconversion to be performed.
Integrating the interpolation filter means that the very high-speed data bus to the DAC
(which may run at 400 MHz or more) need never come off chip, providing significant
improvements in power dissipation and EMI. The AD9777 is designed to mate with
ADI's quadtrature modulator chip AD8346 as a two-chip set, significantly reducing the
number of external components required. Why not integrate the analog mixer? For
performance and testability reasons, it is implemented on a bipolar process. Its
specifications far exceed what is possible in CMOS.
The primary benefit of smart partitioning is the ability to integrate digital functionality
onto high-performance analog circuits and vice-versa. This frees designers to partition
rather than forcing them into a certain arrangement based on the inherent limitations of
their chip's functionality.
Combining functionality in high-performance analog circuitry and high-performance
DSP provides great latitude in partitioning options. This must be combined with a strong
system understanding in order to make the wisest choices.
Why Smart Partitioning is Necessary
A single, dominant force has governed the semiconductor universe over the past 25 years:
the trend toward ever-higher levels of integration. Gordon Moore of Intel even effectively
quantified the slope of this trend, claiming that the level of integration on ICs would
double every 18 months. This has become known as Moore's Law, and has been a
remarkably accurate predictor of the integration trend for semiconductor circuits.
There have been many critical technology advances that have enabled the industry to
keep pace with Moore's prediction. These have included advances to finer and finer
lithography, the ability to handle larger and larger wafers, improvements in chemical
purity, and reductions in defect density. The benefits of marching down this integration
curve have been astounding: exponentially improved processing capability, faster
processing speeds, decreasing costs, and reduced size and power consumption.
While the most notable examples of the integration trend have been seen in memory
circuits (like DRAM) or in microprocessors, the theme of ever-higher levels of
integration pervades virtually every corner of the semiconductor world. The analog world
has been no exception. Over the last 30 years, the state of the art in analog has moved
from operational amplifiers (op-amps) on a single chip, to whole converters, to entire
mixed-signal systems that replace 30 to 50 discrete chips.
Originally, analog integrated circuits were implemented on process technologies that
differed significantly from those used for digital circuits. During the 1980s and 90s, there
was increasing emphasis placed on building analog functions on digital processes,
allowing the analog and digital circuitry to be integrated onto a signal IC—a "mixedsignal" integrated circuit. This has been highly effective, and mixed signal ICs are
pervasive in today's products, from cell phones to digital still cameras. However, the
complexities involved with implementing the analog functions have meant that mixed4.26
DATA CONVERTER PROCESS TECHNOLOGY
4.3 SMART PARTITIONING
signal ICs tend to lag their entirely digital counterparts by at least one lithography
generation. While today's state-of-the art digital circuits are being designed on 0.13-µm
processes, the most advanced mixed-signal circuits are being done on 0.18-µm or
0.25-µm processes.
What's Changing?
As we enter the new millennium, we are starting to see some changes that could have a
significant impact on the seemingly inevitable trend to higher integration. Fundamental
laws of physics may ultimately limit the ability to keep shrinking the lithography. We are
already seeing some significant increases in costs (refer back to Figure 4.15 and Figure
4.16), and in a few generations, we may reach a point where further feature-size
reductions aren't economically practical.
Through most of the 1990s, the analog designers rode the lithography curve that the
digital circuits were pushing. However, late in the decade, there was a significant catch:
with each lithography shrink below 0.5 µm, the maximum allowable supply voltage also
falls. While this was of little consequence to the digital designer (it actually helps to
lower power dissipation), it has enormous significance to the analog designer. Shrinking
supply voltages force the use of smaller signal voltages, making it even more difficult to
preserve the analog signal in the presence of inevitable noise.
Instead of being able to transfer circuit blocks from one lithography generation to the
next, each new lithography sends the analog designer back to the drawing board. The
consequences can be significant. Instead of mixed signal lagging digital by one
lithography generation, this lag is starting to stretch to two to three generations. Some
very high-performance circuits may eventually be impractical (though probably not
impossible) on extremely fine line geometries.
As supply voltages on state-of-the-art digital processes continue to drop, specialized
processes may become more popular for high-performance mixed-signal circuits. Figure
4.20 illustrates that there are other process technology curves that parallel the digital
CMOS curve. These are processes that have been optimized for analog circuits, making
different trade-offs more appropriate to the needs of the analog circuit designers.
In addition to the difficulties of designing analog circuits in smaller geometries, there is
another problem facing "further integration" as the model for future electronic design:
process technology has, in many cases, outpaced design and simulation capability. We
are now able to integrate larger systems than we are able to effectively simulate, analyze
or test.
In the face of these problems, some pundits have predicted that things will go back to the
way they were in the mid-to-late 1980s: digital circuits on one chip, analog circuits on
another, with different process technologies, simulation tools, and designers used for
each. This would essentially constitute a "dis-integration" of the analog from the digital.
While some systems may break down this way, in many cases, the real answer will be
more subtle (and more interesting).
4.27
ANALOG-DIGITAL CONVERSION
100
HIGH VOLTAGE PROCESSES
XFCB
10
SUPPLY
VOLTAGE
(VOLTS)
BiCMOS
1
CMOS
0
0.1
10
1
100
ft (GHz)
Figure 4.20: Range of Semiconductor Processes
There was an evening panel discussion several years ago at one of the major integrated
circuits conferences-the topic was "The Single-chip Cell Phone." The discussion pushed
back and forth about the technical challenges associated with combining the different
chips in a cellular telephone. At one point, a panelist put a picture of a cell phone circuit
board up on the screen. He noted 6 integrated circuits and 300 passive components. "Stop
trying to combine these 6 ICs into one," he said, "and let's do something about all the
passive components!" The lesson from this story is that integration needs to be used in an
intelligent way to reduce the cost, size, and power demands of the overall system, not
simply as an exercise in blindly reducing the IC count.
As we approach the practical limits of integration and lithography, intelligence will need
to be applied as to where and how to integrate. The key is to find optimal points to break
the system into functional blocks. Typically, these will be places that require a minimum
information flow across the boundaries between them, allowing the pin counts (and
therefore cost and size) of the ICs to be kept low. Consideration should also be given to
how the system and ICs will be simulated and analyzed. The partitioning dictated by
these factors may or may not correspond to the boundary between analog and digital.
Analog Devices has been working on chip set partitioning for a number of years, and the
theme of "smart partitioning" has emerged as one of the most significant factors in
optimizing the cost, size, power, and performance for systems that feature both analog
and digital circuits.
Figure 4.21 illustrates an example, taken from a real case, where three chips are cheaper
than one or two. One of the keys is allowing the majority of the digital functionality to be
implemented in the most effective process possible, avoiding the one-to-three generation
4.28
DATA CONVERTER PROCESS TECHNOLOGY
4.3 SMART PARTITIONING
lag normally associated with attempting to integrate everything. Using higherperformance optimized processes for the mixed-signal and analog functions eliminates a
large number of external passive/discrete components, thereby significantly reducing
system size and cost. The mixed-signal front end AD9860/AD9862 as shown in this
application is actually an excellent example for integration of several high-speed
converters onto a single CMOS chip (see also Figure 4.22).
Antenna
Linear Circuitry
Digital Modem
Rx
PHY
AD9860/62
RF / IF
MAC
Beam Former / Equalizer
PLL
ADC
Filter
FFT
OFDM
QAMDecoder
Error
Correction
MAC
RISCProcessor
ADC
RF
Gain
Control
IFHigh IF Low
Timing
Recovery
MAC
Hardware
DAC
Filter
DAC
IFFT
OFDM
QAMCoder
Forward
Error
Correction
Ethernet
Tx
RF / IF
Up / Down Converts
Baseband Signal
to / from RF
Frequencies
AD9860/AD9862
Mixed Signal
Front End
ADCs, DACs
Signal Conditioning
Modulator
Tone Mapper
QAM
OFDM
WCDMA
Diversity Receive
Antenna Combining
in Case of Diversity
Media Access
Control
Channel Sharing
Address Filtering
Security
Figure 4.21: AD9860/AD9862: Broadband Wireless OFDM Modem
Dual 12-Bit, 64MSPS ADCs
Dual 14-Bit,128MSPS DACs
Programmable Filters
and PGAs
Versatile Modulation and
Clock Generation Circuitry
Figure 4.22: AD9860/AD9862 MxFE™ for Broadband Communications
4.29
ANALOG-DIGITAL CONVERSION
In that device, putting dual DACs and dual ADCs on the same piece of silicon guarantees
matching performance, which is important for applications that require IQ modulation
and demodulation. With careful chip design, sufficient isolation between transmit and
receive path has been achieved. Variable gain amplifiers as well as auxiliary ADCs and
DACs have all successfully been integrated. The CMOS process allowed for embedding
complex digital upconversion, interpolation and decimation. This eases digital interface
requirements significantly by reducing the data rate between the digital and mixed-signal
chip. A lower speed interface draws less power and also improves EMI. Radio
components like the mixer, power amplifier and low noise amplifier are not integrated on
this CMOS chip because they achieve better performance when implemented on bipolar
processes.
For many in the IC world, the end of Moore's Law seems unthinkable-ever-shrinking
lithography has come to be viewed as an inalienable right. Nevertheless, the signs
pointing to the end of Moore's law are there for those who will see them, and prudent
designers will adapt to the new realities.
It's worth noting that other industries have faced similar technology limitations and are
still thriving. For example, the ever-upward trend in aircraft speed was remarkably
predictable for forty years, from the Wright brothers to the end of World War II. Yet the
sound barrier posed a technology barrier. While it was possible to fly faster than sound, it
has not proven economically practical to do so for commercial aircraft. Instead, the
aircraft industry has advanced along other dimensions. The electronics industry will do
the same. Smart partitioning will be the industry's way forward.
4.30